NADI 2021: The Second Nuanced Arabic Dialect Identification Shared Task - Computational Approaches to Modeling ...

Page created by Derrick Mitchell
 
CONTINUE READING
NADI 2021: The Second Nuanced Arabic Dialect Identification Shared Task - Computational Approaches to Modeling ...
NADI 2021:
        The Second Nuanced Arabic Dialect Identification Shared Task
          Muhammad Abdul-Mageed, Chiyu Zhang, AbdelRahim Elmadany,
                           Houda Bouamor,† Nizar Habash‡
                The University of British Columbia, Vancouver, Canada
                     †
                       Carnegie Mellon University in Qatar, Qatar
                        ‡
                          New York University Abu Dhabi, UAE
    {muhammad.mageed, a.elmadany}@ubc.ca                    chiyuzh@mail.ubc.ca
               hbouamor@cmu.edu                nizar.habash@nyu.edu

                    Abstract
 We present the findings and results of the
 Second Nuanced Arabic Dialect Identification
 Shared Task (NADI 2021). This Shared Task
 includes four subtasks: country-level Modern
 Standard Arabic (MSA) identification (Subtask
 1.1), country-level dialect identification (Subtask
 1.2), province-level MSA identification (Subtask
 2.1), and province-level sub-dialect identifica-
 tion (Subtask 2.2). The shared task dataset cov-
 ers a total of 100 provinces from 21 Arab coun-       Figure 1: A map of the Arab World showing thr 21
 tries, collected from the Twitter domain. A total     countries and 100 provinces in the NADI 2021 datasets.
 of 53 teams from 23 countries registered to par-      Each country is coded in color different from neighbor-
                                                       ing countries. Provinces within each country are coded
 ticipate in the tasks, thus reflecting the interest
                                                       in a more intense version of the same color as the coun-
 of the community in this area. We received 16         try.
 submissions for Subtask 1.1 from five teams, 27
 submissions for Subtask 1.2 from eight teams,
 12 submissions for Subtask 2.1 from four teams,
                                                       ple from various sub-regions, countries, or even
 and 13 Submissions for subtask 2.2 from four
                                                       provinces within the same country, may be using
 teams.
                                                       Arabic differently.
1   Introduction                                           The Nuanced Arabic Dialect Identification
                                                       (NADI) series of shared tasks aim at furthering
Arabic is the native tongue of ∼ 400 million peo-      the study and analysis of Arabic variants by provid-
ple living the Arab world, a vast geographical re-     ing resources and organizing classification compe-
gion across Africa and Asia. Far from a single         titions under standardized settings. The First Nu-
monolithic language, Arabic has a wide number          anced Arabic Dialect Identification (NADI 2020)
of varieties. In general, Arabic could be classified   Shared Task targeted 21 Arab countries and a to-
into three main categories: (1) Classical Arabic,      tal of 100 provinces across these countries. NADI
the language of the Qur’an and early literature; (2)   2020 consisted of two subtasks: country-level di-
Modern Standard Arabic (MSA), which is usually         alect identification (Subtask 1) and province-level
used in education and formal and pan-Arab media;       detection (Subtask 2). The two subtasks depended
and (3) dialectal Arabic (DA), a collection of geo-    on Twitter data, making it the first shared task to tar-
politically defined variants. Modern day Arabic        get naturally-occurring fine-grained dialectal text
is usually referred to as diglossic with a so-called   at the sub-country level. The Second Nuanced Ara-
‘High’ variety used in formal settings (MSA), and      bic Dialect Identification (NADI 2021) is similar
a ‘Low’ variety used in everyday communication         to NADI 2020 in that it also targets the same 21
(DA). DA, the presumably ‘Low’ variety, is itself      Arab countries and 100 corresponding provinces
a host of variants. For the current work, we focus     and is based on Twitter data. However, NADI 2021
on geography as an axis of variation where peo-        has four subtasks, organized into country level and
province level. For each classification level, we af-         Arabic dialect identification work as further
ford both MSA and DA datasets as Table 1 shows.            sparked by a series of shared tasks offered as part
                                                           of the VarDial workshop. These shared tasks used
       Variety     Country         Province                speech broadcast transcriptions (Malmasi et al.,
       MSA         Subtask 1.1     Subtask 2.1             2016), and integrated acoustic features (Zampieri
       DA          Subtask 1.2     Subtask 2.2             et al., 2017) and phonetic features (Zampieri et al.,
            Table 1: NADI 2021 subtasks.                   2018) extracted from raw audio. Althobaiti (2020)
                                                           is a recent survey of computational work on Arabic
                                                           dialects.
   We provided participants with a new Twitter la-
beled dataset that we collected exclusively for the            The Multi Arabic Dialects Application and Re-
purpose of the shared task. The dataset is publicly        sources (MADAR) project (Bouamor et al., 2018)
available for research.1 A total of 53 teams regis-        introduced finer-grained dialectal data and a lexi-
tered for the shard task, of whom 8 unique teams           con. The MADAR data were used for dialect iden-
ended up submitting their systems for scoring. We          tification at the city level (Salameh et al., 2018;
allowed a maximum of five submissions per team.            Obeid et al., 2019) of 25 Arab cities. An issue with
We received 16 submissions for Subtask 1.1 from            the MADAR data, in the context of DA identifica-
five teams, 27 submissions for Subtask 1.2 from            tion, is that it was commissioned and not naturally
eight teams, 12 submissions for Subtask 2.1 from           occurring. Several larger datasets covering 10-21
four teams, and 13 Submissions for subtask 2.2             countries were also introduced (Mubarak and Dar-
from four teams. We then received seven papers,            wish, 2014; Abdul-Mageed et al., 2018; Zaghouani
all of which we accepted for publication.                  and Charfi, 2018). These datasets come from the
   This paper is organized as follows. We provide          Twitter domain, and hence are naturally-occurring.
a brief overview of the computational linguistic              Several works have also focused on socio-
literature on Arabic dialects in Section 2. We de-         pragmatics meaning exploiting dialectal data.
scribe the two subtasks and dataset in Sections 3          These include sentiment analysis (Abdul-Mageed
and Section 4, respectively. And finally, we intro-        et al., 2014), emotion (Alhuzali et al., 2018),
duce participating teams, shared task results, and         age and gender (Abbes et al., 2020), offen-
a high-level description of submitted systems in           sive language (Mubarak et al., 2020), and sar-
Section 5.                                                 casm (Abu Farha and Magdy, 2020). Concurrent
                                                           with our work, (Abdul-Mageed et al., 2020c) also
2   Related Work                                           describe data and models at country, province, and
As we explained in Section 1, Arabic has three             city levels.
main categories: CA, MSA, and DA. While CA                    The first NADI shared task, NADI 2020 (Abdul-
and MSA have been studied extensively (Harrell,            Mageed et al., 2020b), comprised two subtasks,
1962; Cowell, 1964; Badawi, 1973; Brustad, 2000;           one focusing on 21 Arab countries exploiting Twit-
Holes, 2004), DA has received more attention only          ter data, and another on 100 Arab provinces from
in recent years.                                           the same 21 countries. As is explained in (Abdul-
   One major challenge with studying DA has been           Mageed et al., 2020b), the NADI 2020 datasets
rarity of resources. For this reason, most pioneer-        included a small amount of non-Arabic and also a
ing DA works focused on creating resources, usu-           mixture of MSA and DA. For NADI 2021, we con-
ally for only a small number of regions or coun-           tinue to focus on 21 countries and 100 provinces.
tries (Gadalla et al., 1997; Diab et al., 2010; Al-        However, we breakdown the data into MSA and
Sabbagh and Girju, 2012; Sadat et al., 2014; Smaı̈li       DA for a stronger signal. This also gives us the
et al., 2014; Jarrar et al., 2016; Khalifa et al., 2016;   opportunity to study each of these two main cate-
Al-Twairesh et al., 2018; El-Haj, 2020). A number          gories independently. In other words, in addition to
of works introducing multi-dialectal data sets and         dialect and sub-dialect identification, it allows us
regional level detection models followed (Zaidan           to investigate the extent to which MSA itself can
and Callison-Burch, 2011; Elfardy et al., 2014;            be teased apart at the country and province levels.
Bouamor et al., 2014; Meftouh et al., 2015).               Our hope is that NADI 2021 will support exploring
  1
    The dataset is accessible via our GitHub at: https:    variation in geographical regions that have not been
//github.com/UBC-NLP/nadi.                                 studied before.
3     Task Description                                       While the MADAR shared task (Bouamor et al.,
                                                         2019) involved prediction of a small set of cities,
The NADI shared task consists of four subtasks,          NADI 2020 was the first to propose automatic di-
comprising two levels of classification–country and      alect identification at geographical regions as small
province. Each level of classification is carried out    as provinces. Concurrent with NADI 2020, (Abdul-
for both MSA and DA. We explain the different            Mageed et al., 2020c) introduced the concept of
subtasks across each classification level next.          microdialects, and proposed models for identifying
                                                         language varieties defined at both province and city
3.1    Country-level Classification
                                                         levels. NADI 2021 follows these works, but has
    • Subtask 1.1: Country-level MSA. The goal           one novel aspect: We introduce province-level iden-
      of Subtask 1.1 is to identify country level        tification for MSA and DA independently (i.e., each
      MSA from short written sentences (tweets).         variety is handled in a separate subtask). While
      NADI 2021 Subtask 1.1 is novel since no pre-       province-level sub-dialect identification may be
      vious works focused on teasing apart MSA by        challenging, we hypothesize province-level MSA
      country of origin.                                 might be even more difficult. However, we were
                                                         curious to what extent, if possible at all, a machine
    • Subtask 1.2: Country-level DA. Subtask 1.2         would be successful in teasing apart MSA data at
      is similar to Subtask 1.1, but focuses on iden-    the province-level.
      tifying country level dialect from tweets. Sub-        In addition, similar to NADI 2020, we acknowl-
      task 1.2 is similar to previous works that have    edge that province-level classification is somewhat
      also taken country as their target (Mubarak        related to geolocation prediction exploiting Twit-
      and Darwish, 2014; Abdul-Mageed et al.,            ter data. However, we emphasize that geolocation
      2018; Zaghouani and Charfi, 2018; Bouamor          prediction is performed at the level of users, rather
      et al., 2019; Abdul-Mageed et al., 2020b).         than tweets. This makes our subtasks different
                                                         from geolocation work. Another difference lies in
   We provided labeled data to NADI 2021 partic-         the way we collect our data as we will explain in
ipants with specific training (TRAIN) and devel-         Section 4. Tables 11 and 12 (Appendix A) show
opment (DEV) splits. Each of the 21 labels corre-        the distribution of the 100 province classes in our
sponding to the 21 countries is represented in both      MSA and DA data splits, respectively. Importantly,
TRAIN and DEV. Teams could score their models            for all 4 subtasks, tweets in the TRAIN, DEV and
through an online system (codalab) on the DEV            TEST splits come from disjoint sets.
set before the deadline. We released our TEST
set of unlabeled tweets shortly before the system        3.3     Restrictions and Evaluation Metrics
submission deadline. We then invited participants
                                                         We follow the same general approach to managing
to submit their predictions to the online scoring
                                                         the shared task as our first NADI in 2020. This
system housing the gold TEST set labels. Table 2
                                                         includes providing participating teams with a set of
shows the distribution of the TRAIN, DEV, and
                                                         restrictions that apply to all subtasks, and clear eval-
TEST splits across the 21 countries.
                                                         uation metrics. The purpose of our restrictions is to
3.2    Province-level Classification                     ensure fair comparisons and common experimen-
                                                         tal conditions. In addition, similar to NADI 2020,
    • Subtask 2.1: Province-level MSA. The goal          our data release strategy and our evaluation setup
      of Subtask 2.1 is to identify the specific state   through the CodaLab online platform facilitated
      or province (henceforth, province) from which      the competition management, enhanced timeliness
      an MSA tweet was posted. There are 100             of acquiring results upon system submission, and
      province labels in the data, and provinces         guaranteed ultimate transparency.2
      are unequally distributed among the list of           Once a team registered in the shared task, we
      21 countries.                                      directly provided the registering member with the
                                                         data via a private download link. We provided
    • Subtask 2.2: Province-level DA. Again,
                                                         the data in the form of the actual tweets posted to
      Subtask 2.2 is similar to Subtask 2.1, but the
                                                         the Twitter platform, rather than tweet IDs. This
      goal is identifying the province from which a
                                                            2
      dialectal tweet was posted.                               https://codalab.org/
MSA (Subtasks 1.1 & 2.1)                      DA (Subtasks 1.2 & 2.2)
  Country      Provinces
                           Train     DEV TEST         Total        %     Train     DEV TEST         Total     %
  Algeria          9        1,899      427     439    2,765       8.92    1,809     430      391    2,630    8.48
  Bahrain         1           211       51      51      313       1.01      215      52       52      319    1.03
  Djibouti         1          211       52      51      314       1.01      215      27        7      249    0.80
  Egypt           20        4,220    1,032     989    6,241      20.13    4,283   1,041    1,051    6,375   20.56
  Iraq            13        2,719      671     652    4,042      13.04    2,729     664      664    4,057   13.09
  Jordan          2           422      103     102      627       2.02      429     104      105      638    2.06
  Kuwait           2          422      103     102      627       2.02      429     105      106      640    2.06
  Lebanon         3           633      155     141      929       3.00      644     157      120      921    2.97
  Libya            6        1,266      310     307    1,883       6.07    1,286     314      316    1,916    6.18
  Mauritania       1          211       52      51      314       1.01      215      53       53      321    1.04
  Morocco         4           844      207     205    1,256       4.05      858     207      212    1,277    4.12
  Oman             7        1,477      341     357    2,175       7.02    1,501     355      371    2,227    7.18
  Palestine        2          422      102     102      626       2.02      428     104      105      637    2.05
  Qatar            1          211       52      51      314       1.01      215      52       53      320    1.03
  KSA             10        2,110      510     510    3,130      10.10    2,140     520      522    3,182   10.26
  Somalia          2          346       63     102      511       1.65      172      49       55      276    0.89
  Sudan           1           211       48      51      310       1.00      215      53       53      321    1.04
  Syria            6        1,266      309     306    1,881       6.07    1,287     278      288    1,853    5.98
  Tunisia          4          844      170     176    1,190       3.84      859     173      212    1,244    4.01
  UAE              3          633      154     153      940       3.03      642     157      158      957    3.09
  Yemen           2           422       88     102      612       1.97      429     105      106      640    2.06
  Total           100      21,000    5,000   5,000   31,000       100    21,000   5,000    5,000   31,000    100

      Table 2: Distribution of classes and data splits over our MSA and DA datasets for the four subtasks.
                                                         .

guaranteed comparison between systems exploiting             report their results on the DEV set in their papers.
identical data. For all four subtasks, we provided           We setup four CodaLab competitions for scoring
clear instructions requiring participants not to use         participant systems.3 We will keep the Codalab
any external data. That is, teams were required              competition for each subtask live post competition,
to only use the data we provided to develop their            for researchers who would be interested in train-
systems and no other datasets regardless how these           ing models and evaluating their systems using the
are acquired. For example, we requested that teams           shared task TEST set. For this reason, we will
do not search nor depend on any additional user-             not release labels for the TEST set of any of the
level information such as geolocation. To alleviate          subtasks.
these strict constraints and encourage creative use
of diverse (machine learning) methods in system              4   Shared Task Datasets
development, we provided an unlabeled dataset of
10M tweets in the form of tweet IDs. This dataset            We distributed two Twitter datasets, one in MSA
is in addition to our labeled TRAIN and DEV splits           and another in DA. Each tweet in each of these two
for the four subtasks. To facilitate acquisition of          datasets has two labels, one label for country level
this unlabeled dataset, we also provided a simple            and another label for province level. For example,
script that can be used to collect the tweets. We            for the MSA dataset, the same tweet is assigned
encouraged participants to use these 10M unlabeled           one out of 21 country labels (Subtask 1.1) and one
tweets in any way they wished.                               out of 100 province labels (Subtask 2.1). The same
                                                             applies to DA data, where each tweet is assigned
  For all four subtasks, the official metric is macro-       a country label (Subtask 1.2) and a province label
averaged F 1 score obtained on blind test sets.              (Subtask 2.2). Similar to MSA, the tagset for DA
We also report performance in terms of macro-                data has 21 country labels and 100 province labels.
averaged precision, macro-averaged recall and ac-
                                                                 3
curacy for systems submitted to each of the four                   Links to the CodaLab competitions are as follows:
                                                             Subtask 1.1:      https://competitions.codalab.
subtasks. Each participating team was allowed to             org/competitions/27768, Subtask 1.2: https:
submit up to five runs for each subtask, and only            //competitions.codalab.org/competitions/
the highest scoring run was kept as representing             27769, Subtask 2.1:        https://competitions.
                                                             codalab.org/competitions/27770,                    Sub-
the team. Although official results are based only           task 2.2: https://competitions.codalab.org/
on a blind TEST set, we also asked participants to           competitions/27771.
In addition, as mentioned before, we made avail-            to both our MSA and DA data. That is, we la-
able an unlabeled dataset for optional use in any           bel tweets from each user with the country and
of the four subtasks. We now provide more details           province from which the user consistently posted
about both the labeled and unlabeled data.                  for the whole of the 10 months period. Although
                                                            this method of label assignment is not ideal, it is
4.1   Data Collection                                       still a reasonable approach for easing the bottleneck
Similar to NADI 2020, we used the Twitter API               of data annotation. For both the MSA and DA data,
to crawl data from 100 provinces belonging to 21            across the two levels of classification (i.e., country
Arab countries for 10 months (Jan. to Oct., 2019).4         and province), we randomly sample 21K tweets
Next, we identified users who consistently and ex-          for training (TRAIN), 5K tweets for development
clusively tweeted from a single province during             (DEV), and 5K tweets for testing (TEST). These
the whole 10 month period. We crawled up to                 three splits come from three disjoint sets of users.
3,200 tweets from each of these users. We se-               We distribute data for the four subtasks directly to
lect only tweets assigned the Arabic language tag           participants in the form of actual tweet text. Table 2
(ar) by Twitter. We lightly normalize tweets by             shows the distribution of tweets across the data
removing usernames and hyperlinks, and add white            splits over the 21 countries, for all subtasks. We
space between emojis. Next, we remove retweets              provide the data distribution over the 100 provinces
(i.e., we keep only tweets and replies). Then, we           in Appendix A. More specifically, Table 11 shows
use character-level string matching to remove se-           the province-level distribution of tweets for MSA
quences that have < 3 Arabic tokens.                        (Subtask 2.1) and Table 12 shows the same for DA
                                                            (Subtask 2.2). We provide examples DA tweets
                                                            from a number of countries representing different
                                                            regions in Table 3. For each example in Table 3,
                                                            we list the province it comes from. Similarly, we
                                                            provide example MSA data in Table 4.

                                                            Unlabeled 10M. We shared 10M Arabic tweets
                                                            with participants in the form of tweet IDs. We
                                                            crawled these tweets in 2019. Arabic was identified
                                                            using Twitter language tag (ar). This dataset does
                                                            not have any labels and we call it UNLABELED
                                                            10M. We also included in our data package released
                                                            to participants a simple script to crawl these tweets.
Figure 2: Distribution of tweet length (trimmed at 50)
in words in NADI-2021 labeled data.
                                                            Participants were free to use UNLABELED 10M
                                                            for any of the four subtasks in any way they they
                                                            see fits.5 We now present shared task teams and
   Since the Twitter language tag can be wrong
                                                            results.
sometimes, we apply an effective in-house lan-
guage identification tool on the tweets and replies
                                                            5     Shared Task Teams & Results
to exclude any non-Arabic. This helps us remove
posts in Farsi (fa) and Persian (ps) which Twitter          5.1    Our Baseline Systems
wrongly assigned an Arabic language tag. Finally,
to tease apart MSA from DA, we use the dialect-             We provide two simple baselines, Baseline I and
MSA model introduced in Abdul-Mageed et al.                 Baseline II, for each of the four subtasks. Base-
(2020a) (acc= 89.1%, F1= 88.6%).                            line I is based on the majority class in the TRAIN
                                                            data for each subtask. It performs at F1 = 1.57%
4.2   Data Sets                                             and accuracy = 19.78% for Subtask 1.1, F1 =
                                                            1.65% and accuracy = 21.02% for Subtask 1.2,
To assign labels for the different subtasks, we use
                                                            F1 = 0.02% and accuracy = 1.02% for Subtask
user location as a proxy for language variety labels
at both country and province levels. This applies              5
                                                                 Datasets for all the subtasks and UNLABELED 10M are
                                                            available at https://github.com/UBC-NLP/nadi.
  4
    Although we tried, we could not collect data from Co-   More information about the data format can be found in the
moros to cover all 22 Arab countries.                       accompanying README file.
Country      Province                                                                  Tweet
      Algeria      Bouira                                                             úæJ Kñ» ø YK úm.' ... éJ¯ ùªJ.K ú»@P ÈAm
                   Khenchela                                                      !, PQK @QË@ ñƒY®K úÍ ¼ð XAë ÕæË@ ñJ.Ê®K ÈAm
                   Oran                                                                                              @ñJm'. P éK QÓ àA®«P ¼@P
      Egypt        Alexandria                                                                 AJKAJk ø YªK PY® Jë AÓ ø P Ó                 .
                   Minya                                                                                      
                                                                         úÍñêÊJ®K ½ÊJj.ë úkA“ ù®J.Jë ø Ò’Ó É¿ AK@ ‘.        Q
                   Sohag                                                                       éJËñË@ ø P úæj‚®K Yg  YJªÓ AK@
      KSA          Ar-Riyad                                                                      !! €AÖ   Þ àñJ.ªÊK àñ‚Êm.' áK YªK ð
                   Ash-Sharqiyah                             Ñê®K ÑêÖßQk úæ.K àñÒê®K AÓ éK ð@Q® ¯ ÈAg. P ÑêÖÞ @ ÈCmÌ '@ áK. AK
                   Tabuk                                                            éJªJJ.£ èA¯ð ø Yë èP@PñË@ ÉgX  @ I.J£
      Morocco      Marrakech-Tensift-Al-Haouz                                  ÈðXAJK. H. Qå @ éJ .’ªË@ €Q»
                                                                    hAKP@ àA‚«                                                 Bð l'. P ú¯AÓ
                   Meknes-Tafilalet                                           . AmÌ áÊ¿@YÓ@ p úG P ñ®«@ p@X @ qJK@ àð@X@QÓ
                                                                        !!!! Ik                               .
                   Souss-Massa-Draa                                                 ¬@QK. áJJªJ‚ Ó AÒë úæk              ¨A¿ ÑîDJ ‚ éJK @
                                                                                                              Yg C“@         
      Oman         Ash-Sharqiyah                                                                       .lÌ. ‡ÊªK                       úæK@ i.ʒm'
                   Dhofar                                                                                éJÊ« ᮪K úæ.JÓ , à@Q¯ ùªÖÞ @
                   Musandam                              AîE PAm.' é
Country     Province                                                                                Tweet
 Algeria     Biskra                                                                                                        ÉJj‚Ó   iJ“@.. ÈCg ‡k           ¼YJ«
                                                                                                                                              .          .
                                                                                         AK. AJ« ð ... AK. A¢k H Qêk. ZA®ÊË@       éJ J« úΫ CK ñ£ ½KQ¢JK@
                                                                                                                                           .
             Oran
                                                                      ý«ð@ ... ú¾K. @ à@ YK P@ úæK@ ... @QJ» ½KQ.g@ .. A¯A J« Q« HXP@                 AÓ ð ..
             Ouargla
                                                                                                                               
                                                                                                  ½K QK ñK. @ éÊ KAªË éJ m' ­Ë A¯ é£A‚   Ë@ ñë úæƒ ÉÔg @
                                                                                                                                                   .                 .
 Egypt       Faiyum                                                                                                      Ðð QêÓ €PA¯ ½«ñK. P á« ÉgQK àB@
             Minya
                                                        á« É¿ áÓð éÓAëð
                                                       éÓB                                      àA¢Jƒ É¿ áÓ éÓA          JË@ é
Team                         F1             Acc       Precision       Recall
                           CairoSquad               22.38(1)        35.72(1)     31.56(1)      20.66(1)
                           Phonemer                 21.79(2)        32.46(3)     30.03(3)      19.95(2)
                           CS-UM6P                  21.48(3)        33.74(2)     30.72(2)      19.70(3)
                           Speech Translation       14.87(4)        24.32(4)     18.95(4)      13.85(4)
                           Our Baseline II          14.15           24.76        20.01         13.21
                           NAYEL                    12.99(5)        23.24(5)     15.09(5)      12.46(5)
                           Our Baseline I            1.57           19.78         0.94          4.76

Table 6: Results for Subtask 1.1 (country-level MSA). The numbers in parentheses are the ranks. The table is
sorted on the macro − F 1 score, the official metric.

                                   Team                 F1             Acc       Precision       Recall
                           CairoSquad               32.26(1)        51.66(1)     36.03(1)      31.09(1)
                           CS-UM6P                  30.64(2)        49.50(2)     32.91(2)      30.34(2)
                           Phonemer                 24.29(4)        44.14(3)     30.24(3)      23.70(4)
                           Speech Translation       21.49(5)        40.54(5)     26.75(5)      20.36(6)
                           Arizona                  21.37(6)        40.46(6)     26.32(6)      20.78(5)
                           AraDial MJ               18.94(7)        35.94(8)     21.58(8)      18.28(7)
                           NAYEL                    18.72(8)        37.16(7)     21.61(7)      18.12(8)
                           Our Baseline II          18.02           33.04        18.69         17.88
                           Our Baseline I            1.65           21.02         1.00          4.76

                                   Table 7: Results for Subtask 1.2 (province-level MSA)

                                     Team               F1            Acc      Precision      Recall
                               CairoSquad           6.43(1)         6.66(1)    7.11(1)       6.71(1)
                               Phonemer             5.49(2)         6.00(2)    6.17(2)       6.07(2)
                               CS-UM6P              5.35(3)         5.72(3)    5.71(3)       5.75(3)
                               NAYEL                3.51(4)         3.38(4)    4.09(4)       3.45(4)
                               Our Baseline II      3.39            3.48       3.68          3.49
                               Our Baseline I       0.02            1.02       0.01          1.00

                                    Table 8: Results for Subtask 2.1 (country-level DA).

                                     Team               F1            Acc      Precision      Recall
                               CairoSquad           8.60(1)         9.46(1)    9.07(1)       9.33(1)
                               CS-UM6P              7.32(2)         7.92(2)    7.73(2)       7.95(2)
                               NAYEL                4.55(3)         4.80(3)    4.71(3)       4.55(4)
                               Phonemer             4.37(4)         5.32(4)    4.49(4)       5.19(3)
                               Our Baseline II      4.08            4.18       4.54          4.22
                               Our Baseline I       0.02            1.06       0.01          1.00

                                   Table 9: Results for Subtask 2.2 (province-level DA).

mance with 8.60%.7                                                     and CS-UM6P developed their system utilizing
                                                                       MARBERT (Abdul-Mageed et al., 2020a), a pre-
5.4    General Description of Submitted                                trained Transformer language model tailored to
       Systems                                                         Arabic dialects and the domain of social media.
In Table 10, we provide a high-level descrip-                          Team Phonemer utilized AraBERT (Antoun et al.,
tion of the systems submitted to each subtask.                         2020a) and AraELECTRA (Antoun et al., 2020b).
For each team, we list their best score of each                        Team CairSquad apply adapter modules (Houlsby
subtask, the features employed, and the meth-                          et al., 2019) and vertical attention to MARBERT
ods adopted/developed. As can be seen from                             fine-tuning. CS-UM6P fine-tuned MARBERT on
the table, the majority of the top teams have                          country-level and province-level jointly by multi-
used Transformers. Specifically, team CairoSquad                       task learning. The rest of participating teams have
    7
                                                                       either used a type of neural networks other than
      The full sets of results for Subtask 1.1, 1.2, 2.1, and 2.2
are in Tables 13, 14, 15 and 15, respectively, in Appendix A.          Transformers or resorted to linear machine learning
Features                                                                    Techniques

                                                                    Word embeds

                                                                                                   Classical ML

                                                                                                                                  Transformer
                                                                                                                  Neural nets

                                                                                                                                                                       Semi-super
                                                    Linguistics

                                                                                                                                                           Multitask
                                                                                                                                                Ensemble
                                                                                        Sampling
                                           TF-IDF
                                  N-gram

                                                                                  PMI
                            F1
               Team

                                                                  SUBTASK 1.1
           CairoSquad     22.38
           Phonemer       21.79
           CS-UM6P        21.48
           Speech Trans   14.87
           NAYEL          12.99
                                                                  SUBTASK 1.2
           CairoSquad     32.26
           CS-UM6P        30.64
           Phonemer       24.29
           Speech Trans   21.49
           Arizona        21.37
           AraDial MJ     18.94
           NAYEL          18.72
                                                                  SUBTASK 2.1
           CairoSquad      6.43
           Phonemer        5.49
           CS-UM6P         5.35
           NAYEL           3.51
                                                                  SUBTASK 2.2
           CairoSquad      8.60
           CS-UM6P         7.32
           NAYEL           4.55
           Phonemer        4.37

Table 10: Summary of approaches used by participating teams. PMI: poinwise mutual information. Classical ML
refers to any non-neural machine learning methods such as naive Bayes and support vector machines. The term
“neural nets” refers to any model based on neural networks (e.g., FFNN, RNN, and CNN) except Transformer
models. Transformer refers to neural networks based on a Transformer architecture such as BERT. The table is
sorted by official metric , macro − F1 . We only list teams that submitted a description paper. “Semi-super”
indicates that the model is trained with semi-supervised learning.

models, usually with some form of ensembling.                                     a continued interest in the community and calls for
                                                                                  further work in this area.
6   Conclusion and Future Work                                                       In the future, we plan to host a third iteration of
                                                                                  the NADI shared task that will use new datasets and
We presented the findings and results of the NADI                                 encourage novel solutions to the set of problems
2021 shared task. We described our datasets across                                introduced in NADI 2021. As results show all the
the four subtasks and the logistics of running the                                fours subtasks remain very challenging, and we
shared task. We also provided a panoramic descrip-                                hope that encouraging further solutions will help
tion of the methods used by all participating teams.                              advance work in this area.
The results show that distinguishing the language
variety of short texts based on small geographical                                Acknowledgments
regions of origin is possible, yet challenging. The
total number of submissions during official evalua-                               We gratefully acknowledge the support of the Nat-
tion (n=68 submissions from 8 unique teams), as                                   ural Sciences and Engineering Research Council
well as the number of teams who registered and                                    of Canada, the Social Sciences Research Council
acquired our datasets (n=53 unique teams) reflects                                of Canada, Compute Canada, and UBC Sockeye.
References                                                    Modeling of People’s Opinions, Personality, and
                                                              Emotions in Social Media, pages 25–35.
Ines Abbes, Wajdi Zaghouani, Omaima El-Hardlo, and
   Faten Ashour. 2020. Daict: A dialectal Arabic irony      Badr AlKhamiss, Mohamed Gabr, Muhammed El-
   corpus extracted from twitter. In Proceedings of           Nokrashy, and Khaled Essam. 2021. Adapting
  The 12th Language Resources and Evaluation Con-             MARBERT for Improved Arabic Dialect Identifica-
   ference, pages 6265–6271.                                  tion: Submission to the NADI 2021 Shared Task. In
                                                              Proceedings of the Sixth Arabic Natural Language
Muhammad Abdul-Mageed, Hassan Alhuzali, and Mo-               Processing Workshop (WANLP 2021).
 hamed Elaraby. 2018. You tweet what you speak: A
 city-level dataset of Arabic dialects. In Proceedings      Maha J Althobaiti. 2020. Automatic Arabic dialect
 of the Language Resources and Evaluation Confer-            identification systems for written texts: A survey.
 ence (LREC), Miyazaki, Japan.                               arXiv preprint arXiv:2009.12622.
Muhammad Abdul-Mageed, Mona Diab, and Sandra                Maha J. Althobaiti. 2021. Country-level Arabic Di-
 Kübler. 2014. Samar: Subjectivity and sentiment            alect Identification Using Small Datasets with In-
 analysis for Arabic social media. Computer Speech           tegrated Machine Learning Techniques and Deep
 and Language, 28(1):20–37.                                  Learning Models. In Proceedings of the Sixth
                                                             Arabic Natural Language Processing Workshop
Muhammad Abdul-Mageed, AbdelRahim Elmadany,                  (WANLP 2021).
 and El Moatez Billah Nagoudi. 2020a. ARBERT
 & MARBERT: Deep Bidirectional Transformers for             Wissam Antoun, Fady Baly, and Hazem Hajj. 2020a.
 Arabic. arXiv preprint arXiv:2010.04900.                     Arabert: Transformer-based model for arabic lan-
                                                              guage understanding. In Proceedings of the 4th
Muhammad Abdul-Mageed, Chiyu Zhang, Houda                    Workshop on Open-Source Arabic Corpora and Pro-
 Bouamor, and Nizar Habash. 2020b. NADI 2020:                 cessing Tools, with a Shared Task on Offensive Lan-
 The First Nuanced Arabic Dialect Identification              guage Detection, pages 9–15.
 Shared Task. In Proceedings of the Fifth Arabic
 Natural Language Processing Workshop (WANLP                Wissam Antoun, Fady Baly, and Hazem Hajj. 2020b.
 2020), pages 97–110, Barcelona, Spain.                       AraELECTRA: pre-training text discriminators for
                                                              arabic language understanding.   arXiv preprint
Muhammad Abdul-Mageed, Chiyu Zhang, Ab-
                                                              arXiv:2012.15516.
 delRahim Elmadany, and Lyle Ungar. 2020c.
 Micro-dialect identification in diaglossic and code-       MS Badawi. 1973. Levels of contemporary Arabic in
 switched environments. In Proceedings of the                Egypt. Cairo: Dâr al Ma’ârif.
 2020 Conference on Empirical Methods in Natural
 Language Processing (EMNLP), pages 5855–5876.              Houda Bouamor, Nizar Habash, and Kemal Oflazer.
                                                              2014. A multidialectal parallel corpus of Arabic. In
Ibrahim Abu Farha and Walid Magdy. 2020. From Ara-            Proceedings of the Language Resources and Evalu-
   bic Sentiment Analysis to Sarcasm Detection: The           ation Conference (LREC), Reykjavik, Iceland.
   ArSarcasm Dataset. In Proceedings of the 4th Work-
   shop on Open-Source Arabic Corpora and Process-          Houda Bouamor, Nizar Habash, Mohammad Salameh,
   ing Tools, with a Shared Task on Offensive Language        Wajdi Zaghouani, Owen Rambow, Dana Abdul-
   Detection, pages 32–39.                                    rahim, Ossama Obeid, Salam Khalifa, Fadhl Eryani,
                                                              Alexander Erdmann, and Kemal Oflazer. 2018. The
Rania Al-Sabbagh and Roxana Girju. 2012. YADAC:               MADAR Arabic Dialect Corpus and Lexicon. In
  Yet another Dialectal Arabic Corpus. In Proceed-            Proceedings of the Language Resources and Eval-
  ings of the Eighth International Conference on              uation Conference (LREC), Miyazaki, Japan.
  Language Resources and Evaluation (LREC-2012),
  pages 2882–2889.                                          Houda Bouamor, Sabit Hassan, and Nizar Habash.
                                                              2019. The MADAR shared task on Arabic fine-
Nora Al-Twairesh, Rawan Al-Matham, Nora Madi,                 grained dialect identification. In Proceedings of the
  Nada Almugren, Al-Hanouf Al-Aljmi, Shahad Al-               Fourth Arabic Natural Language Processing Work-
  shalan, Raghad Alshalan, Nafla Alrumayyan, Shams            shop, pages 199–207.
  Al-Manea, Sumayah Bawazeer, Nourah Al-Mutlaq,
  Nada Almanea, Waad Bin Huwaymil, Dalal Alqu-              Kristen Brustad. 2000. The Syntax of Spoken Arabic:
  sair, Reem Alotaibi, Suha Al-Senaydi, and Abeer             A Comparative Study of Moroccan, Egyptian, Syrian,
  Alfutamani. 2018. SUAR: Towards building a cor-             and Kuwaiti Dialects. Georgetown University Press.
  pus for the Saudi dialect. In Proceedings of the Inter-
  national Conference on Arabic Computational Lin-          Mark W. Cowell. 1964. A Reference Grammar of Syr-
  guistics (ACLing).                                         ian Arabic. Georgetown University Press, Washing-
                                                             ton, D.C.
Hassan Alhuzali, Muhammad Abdul-Mageed, and
  Lyle Ungar. 2018. Enabling deep learning of emo-          Mona Diab, Nizar Habash, Owen Rambow, Mohamed
  tion with first-person seed expressions. In Pro-           Altantawy, and Yassine Benajiba. 2010. COLABA:
  ceedings of the Second Workshop on Computational           Arabic dialect annotation and processing. In LREC
workshop on Semitic language processing, pages             Dialect Identification based on Weighted Concate-
  66–74.                                                     nation of TF-IDF Transformers. In Proceedings of
                                                             the Sixth Arabic Natural Language Processing Work-
Mahmoud El-Haj. 2020. Habibi - a multi dialect multi         shop (WANLP 2021).
 national Arabic song lyrics corpus. In Proceed-
 ings of the 12th Language Resources and Evaluation        Shervin Malmasi, Marcos Zampieri, Nikola Ljubešić,
 Conference, pages 1318–1326, Marseille, France.             Preslav Nakov, Ahmed Ali, and Jörg Tiedemann.
                                                             2016. Discriminating between similar languages
Abdellah El Mekki, Abdelkader El Mahdaouy, Ka-               and arabic dialect identification: A report on the
  bil Essefar, Nabil El Mamoun, Ismail Berrada, and          third dsl shared task. In Proceedings of the third
  Ahmed Khoumsi. 2021. CS-UM6P @ NADI’2021:                  workshop on NLP for similar languages, varieties
  BERT-based Multi-Task Model for Country and                and dialects (VarDial3), pages 1–14.
  Province Level MSA and Dialectal Arabic Identifi-
  cation. In Proceedings of the Sixth Arabic Natural       Karima Meftouh, Salima Harrat, Salma Jamoussi,
  Language Processing Workshop (WANLP 2021).                 Mourad Abbas, and Kamel Smaili. 2015. Machine
                                                             translation experiments on padic: A parallel Arabic
Heba Elfardy, Mohamed Al-Badrashiny, and Mona
                                                             dialect corpus. In Proceedings of the Pacific Asia
  Diab. 2014. Aida: Identifying code switching in
                                                             Conference on Language, Information and Compu-
  informal Arabic text. In Proceedings of the Con-
                                                             tation.
  ference on Empirical Methods in Natural Language
  Processing (EMNLP), pages 94–101, Doha, Qatar.
                                                           Hamdy Mubarak and Kareem Darwish. 2014. Using
Hassan Gadalla, Hanaa Kilany, Howaida Arram,                 Twitter to collect a multi-dialectal corpus of Arabic.
  Ashraf Yacoub, Alaa El-Habashi, Amr Shalaby,               In Proceedings of the Workshop for Arabic Natural
  Krisjanis Karins, Everett Rowson, Robert MacIn-            Language Processing (WANLP), Doha, Qatar.
  tyre, Paul Kingsbury, David Graff, and Cynthia
  McLemore. 1997. CALLHOME Egyptian Arabic                 Hamdy Mubarak, Kareem Darwish, Walid Magdy,
  transcripts LDC97T19. Web Download. Philadel-              Tamer Elsayed, and Hend Al-Khalifa. 2020.
  phia: Linguistic Data Consortium.                          Overview of osact4 Arabic offensive language de-
                                                             tection shared task. In Proceedings of the 4th Work-
R.S. Harrell. 1962. A Short Reference Grammar of Mo-         shop on Open-Source Arabic Corpora and Process-
  roccan Arabic: With Audio CD. Georgetown clas-             ing Tools, with a Shared Task on Offensive Language
  sics in Arabic language and linguistics. Georgetown        Detection, pages 48–52.
  University Press.
                                                           Hamada Nayel, Ahmed Hassan, Mahmoud Sobhi, and
Clive Holes. 2004. Modern Arabic: Structures, Func-          Ahmed El-Sawy. 2021. Data-Driven Approach for
   tions, and Varieties. Georgetown Classics in Ara-         Arabic Dialect Identification. In Proceedings of the
   bic Language and Linguistics. Georgetown Univer-          Sixth Arabic Natural Language Processing Work-
   sity Press.                                               shop (WANLP 2021).
Neil Houlsby, Andrei Giurgiu, Stanislaw Jastrzebski,       Ossama Obeid, Mohammad Salameh, Houda Bouamor,
  Bruna Morrone, Quentin De Laroussilhe, Andrea              and Nizar Habash. 2019. ADIDA: Automatic di-
  Gesmundo, Mona Attariyan, and Sylvain Gelly.               alect identification for Arabic. In Proceedings of
  2019. Parameter-efficient transfer learning for nlp.       the 2019 Conference of the North American Chap-
  In International Conference on Machine Learning,           ter of the Association for Computational Linguistics
  pages 2790–2799. PMLR.                                     (Demonstrations), pages 6–11, Minneapolis, Min-
                                                             nesota. Association for Computational Linguistics.
Elsayed Issa. 2021. Country-level Arabic dialect identi-
   fication using RNNs with and without linguistic fea-
                                                           Fatiha Sadat, Farnazeh Kazemi, and Atefeh Farzindar.
   tures. In Proceedings of the Sixth Arabic Natural
                                                             2014. Automatic identification of Arabic language
  Language Processing Workshop (WANLP 2021).
                                                             varieties and dialects in social media. Proceedings
Mustafa Jarrar, Nizar Habash, Faeq Alrimawi, Diyam           of SocialNLP, page 22.
 Akra, and Nasser Zalmout. 2016. Curras: an anno-
 tated corpus for the Palestinian Arabic dialect. Lan-     Mohammad Salameh, Houda Bouamor, and Nizar
 guage Resources and Evaluation, pages 1–31.                Habash. 2018. Fine-grained Arabic dialect identi-
                                                            fication. In Proceedings of the International Con-
Salam Khalifa, Nizar Habash, Dana Abdulrahim, and           ference on Computational Linguistics (COLING),
  Sara Hassan. 2016. A Large Scale Corpus of                pages 1332–1344, Santa Fe, New Mexico, USA.
  Gulf Arabic. In Proceedings of the Language Re-
  sources and Evaluation Conference (LREC), Por-           Kamel Smaı̈li, Mourad Abbas, Karima Meftouh, and
  torož, Slovenia.                                          Salima Harrat. 2014. Building resources for Alge-
                                                             rian Arabic dialects. In Proceedings of the Confer-
Mohamed Lichouri, Mourad Abbas, Khaled Lounnas,              ence of the International Speech Communication As-
 Besma Benaziz, and Aicha Zitouni. 2021. Arabic              sociation (Interspeech).
Anshul Wadhawan. 2021. Dialect Identification in Nu-
  anced Arabic Tweets Using Farasa Segmentation
  and AraBERT. In Proceedings of the Sixth Arabic
  Natural Language Processing Workshop (WANLP
  2021).
Wajdi Zaghouani and Anis Charfi. 2018. ArapTweet:
 A Large Multi-Dialect Twitter Corpus for Gender,
 Age and Language Variety Identification. In Pro-
  ceedings of the Language Resources and Evaluation
 Conference (LREC), Miyazaki, Japan.
Omar F Zaidan and Chris Callison-Burch. 2011. The
 Arabic online commentary dataset: an annotated
 dataset of informal Arabic with high dialectal con-
 tent. In Proceedings of the 49th Annual Meeting of
 the Association for Computational Linguistics: Hu-
 man Language Technologies: short papers-Volume
 2, pages 37–41. Association for Computational Lin-
 guistics.

Marcos Zampieri, Shervin Malmasi, Nikola Ljubešić,
 Preslav Nakov, Ahmed Ali, Jörg Tiedemann, Yves
 Scherrer, and Noëmi Aepli. 2017. Findings of the
 vardial evaluation campaign 2017.

Marcos Zampieri, Shervin Malmasi, Preslav Nakov,
 Ahmed Ali, Suwon Shon, James Glass, Yves Scher-
 rer, Tanja Samardzic, Nikola Ljubešić, Jörg Tiede-
 mann, et al. 2018. Language identification and mor-
 phosyntactic tagging: The second vardial evaluation
 campaign. In Proceedings of the Fifth Workshop on
 NLP for Similar Languages, Varieties and Dialects
 (VarDial 2018), pages 1–17.

Appendices
A    Data
We provide the distribution Distribution of the
NADI 2021 MSA data over provinces, by country
(Subtask 2.1), across our our data splits in Table 11.
Similarly, Table 12 shows the distribution of the DA
data over provinces for all countries (Subtask 2.2)
in our data splits.

B    Shared Task Teams & Results
We provide full results for all four subtasks. Ta-
ble 13 shows full results for Subtask 1.1, Table 14
for Subtask 1.2, Table 15 for Subtask 2.1, and Ta-
ble 16 for Subtask 2.2.
Province Name     TRAIN DEV TEST Province Name               TRAIN DEV TEST
          ae Abu-Dhabi       211   51  51 kw Jahra                      211   52  51
          ae Dubai           211   52  51 lb Akkar                      211   52  39
          ae Ras-Al-Khaymah 211    51  51 lb North-Lebanon              211   51  51
          bh Capital         211   51  51 lb South-Lebanon              211   52  51
          dj Djibouti        211   52  51 ly Al-Butnan                  211   52  51
          dz Batna           211   52  51 ly Al-Jabal-al-Akhdar         211   52  52
          dz Biskra          211   52  51 ly Benghazi                   211   51  51
          dz Bouira          211   12  51 ly Darnah                     211   52  51
          dz Béchar         211   52  31 ly Misrata                    211   52  51
          dz Constantine     211   51  51 ly Tripoli                    211   51  51
          dz El-Oued         211   52  51 ma Marrakech-Tensift-Al-Haouz 211   51  51
          dz Khenchela       211   52  51 ma Meknes-Tafilalet           211   52  52
          dz Oran            211   52  51 ma Souss-Massa-Draa           211   52  51
          dz Ouargla         211   52  51 ma Tanger-Tetouan             211   52  51
          eg Alexandria      211   51  51 mr Nouakchott                 211   52  51
          eg Aswan           211   52  51 om Ad-Dakhiliyah              211   51  51
          eg Asyut           211   52  51 om Ad-Dhahirah                211   32  51
          eg Beheira         211   52  51 om Al-Batnah                  211   51  51
          eg Beni-Suef       211   52  51 om Ash-Sharqiyah              211   51  51
          eg Dakahlia        211   51  51 om Dhofar                     211   52  51
          eg Faiyum          211   52  51 om Musandam                   211   52  51
          eg Gharbia         211   52  51 om Muscat                     211   52  51
          eg Ismailia        211   52  51 ps Gaza-Strip                 211   51  51
          eg Kafr-el-Sheikh  211   52  20 ps West-Bank                  211   51  51
          eg Luxor           211   52  51 qa Ar-Rayyan                  211   52  51
          eg Minya           211   51  51 sa Al-Madinah                 211   51  51
          eg Monufia         211   52  51 sa Al-Quassim                 211   51  51
          eg North-Sinai     211   52  51 sa Ar-Riyad                   211   51  51
          eg Port-Said       211   51  51 sa Ash-Sharqiyah              211   51  51
          eg Qena            211   51  51 sa Asir                       211   51  51
          eg Red-Sea         211   52  51 sa Ha’il                      211   51  51
          eg Sohag           211   51  51 sa Jizan                      211   51  51
          eg South-Sinai     211   51  51 sa Makkah                     211   51  51
          eg Suez            211   51  51 sa Najran                     211   51  51
          iq Al-Anbar        211   51  51 sa Tabuk                      211   51  51
          iq Al-Muthannia    211   52  51 sd Khartoum                   211   48  51
          iq An-Najaf        211   51  51 so Banaadir                   211   52  51
          iq Arbil           211   52  51 so Woqooyi-Galbeed            135   11  51
          iq As-Sulaymaniyah 187   52  51 sy Aleppo                     211   51  51
          iq Babil           211   52  51 sy As-Suwayda                 211   51  51
          iq Baghdad         211   51  51 sy Damascus-City              211   51  51
          iq Basra           211   51  51 sy Hama                       211   52  51
          iq Dihok           211   52  51 sy Hims                       211   52  51
          iq Karbala         211   52  40 sy Lattakia                   211   52  51
          iq Kirkuk          211   52  51 tn Ariana                     211   51  51
          iq Ninawa          211   52  51 tn Bizerte                    211   15  51
          iq Wasit           211   51  51 tn Mahdia                     211   52  23
          jo Aqaba           211   52  51 tn Sfax                       211   52  51
          jo Zarqa           211   51  51 ye Aden                       211   51  51
          kw Hawalli         211   51  51 ye Ibb                        211   37  51

Table 11: Distribution of the NADI 2021 MSA data over provinces, by country, across our TRAIN, DEV, and
TEST splits (Subtask 2.1).
Province Name        TRAIN    DEV    TEST    Province Name                   TRAIN    DEV    TEST
    ae Abu-Dhabi          214      52     52     kw Jahra                         215      53     53
    ae Dubai              214      53     53     lb Akkar                         215      53     14
    ae Ras-Al-Khaymah     214      52     53     lb North-Lebanon                 215      52     53
    bh Capital            215      52     52     lb South-Lebanon                 214      52     53
    dj Djibouti           215      27      7     ly Al-Butnan                     214      52     53
    dz Batna              215      34     10     ly Al-Jabal-al-Akhdar            215      53     53
    dz Biskra             215      53     53     ly Benghazi                      214      52     52
    dz Bouira             215      26     53     ly Darnah                        215      53     53
    dz Béchar            215      53     11     ly Misrata                       214      52     53
    dz Constantine        215      52     53     ly Tripoli                       214      52     52
    dz El-Oued            215      53     52     ma Marrakech-Tensift-Al-Haouz    214      52     53
    dz Khenchela           89      53     53     ma Meknes-Tafilalet              215      50     53
    dz Oran               215      53     53     ma Souss-Massa-Draa              215      53     53
    dz Ouargla            215      53     53     ma Tanger-Tetouan                214      52     53
    eg Alexandria         214      52     52     mr Nouakchott                    215      53     53
    eg Aswan              214      52     52     om Ad-Dakhiliyah                 214      52     53
    eg Asyut              214      53     53     om Ad-Dhahirah                   215      40     53
    eg Beheira            214      52     52     om Al-Batnah                     214      52     53
    eg Beni-Suef          214      52     52     om Ash-Sharqiyah                 214      52     53
    eg Dakahlia           214      52     52     om Dhofar                        214      53     53
    eg Faiyum             214      52     53     om Musandam                      215      53     53
    eg Gharbia            214      52     53     om Muscat                        215      53     53
    eg Ismailia           214      52     53     ps Gaza-Strip                    214      52     52
    eg Kafr-el-Sheikh     215      52     53     ps West-Bank                     214      52     53
    eg Luxor              214      52     52     qa Ar-Rayyan                     215      52     53
    eg Minya              214      52     53     sa Al-Madinah                    214      52     52
    eg Monufia            215      52     53     sa Al-Quassim                    214      52     52
    eg North-Sinai        215      52     53     sa Ar-Riyad                      214      52     52
    eg Port-Said          214      52     52     sa Ash-Sharqiyah                 214      52     52
    eg Qena               214      52     53     sa Asir                          214      52     52
    eg Red-Sea            214      52     53     sa Ha’il                         214      52     52
    eg Sohag              214      52     52     sa Jizan                         214      52     53
    eg South-Sinai        214      52     53     sa Makkah                        214      52     52
    eg Suez               214      52     52     sa Najran                        214      52     53
    iq Al-Anbar           214      52     52     sa Tabuk                         214      52     52
    iq Al-Muthannia       215      53     53     sd Khartoum                      215      53     53
    iq An-Najaf           215      53     53     so Banaadir                      136      40      2
    iq Arbil              215      53     53     so Woqooyi-Galbeed                36       9     53
    iq As-Sulaymaniyah    153      32     53     sy Aleppo                        215      52     23
    iq Babil              215      53     53     sy As-Suwayda                    214      53     53
    iq Baghdad            214      52     52     sy Damascus-City                 214      52     53
    iq Basra              214      52     53     sy Hama                          215      53     53
    iq Dihok              215      53     30     sy Hims                          214      53     53
    iq Karbala            215      53     53     sy Lattakia                      215      15     53
    iq Kirkuk             215      53     53     tn Ariana                        214      52     53
    iq Ninawa             215      53     53     tn Bizerte                       215      16     53
    iq Wasit              214      52     53     tn Mahdia                        215      52     53
    jo Aqaba              215      52     53     tn Sfax                          215      53     53
    jo Zarqa              214      52     52     ye Aden                          214      52     53
    kw Hawalli            214      52     53     ye Ibb                           215      53     53

Table 12: Distribution of the NADI 2021 DA data over provinces, by country, across our TRAIN, DEV, and TEST
splits (Subtask 2.2).
Team              F1          Acc      Precision     Recall
                       CairoSquad           22.38(1)    35.72(1)    31.56(3)     20.66(1)
                       CairoSquad           21.97(2)    34.90(2)    30.01(7)     20.15(2)
                       Phonemer             21.79(3)    32.46(6)    30.03(6)     19.95(4)
                       Phonemer             21.66(4)    31.70(7)    28.46(8)     20.01(3)
                       CS-UM6P              21.48(5)    33.74(4)    30.72(5)     19.70(5)
                       CS-UM6P              20.91(6)    33.84(3)    31.16(4)     19.09(6)
                       Phonemer             20.78(7)    32.96(5)    37.69(1)     18.42(8)
                       CS-UM6P              19.80(8)    31.68(8)    26.69(9)     19.04(7)
                       Speech Translation   14.87(9)    24.32(11)   18.95(14)    13.85(9)
                       Speech Translation   14.50(10)   24.06(12)   20.24(12)    13.24(10)
                       Speech Translation   14.48(11)   24.88(9)    22.88(10)    13.17(11)
                       NAYEL                12.99(12)   23.24(14)   15.09(15)    12.46(12)
                       NAYEL                11.84(13)   23.74(13)   19.42(13)    10.92(13)
                       NAYEL                10.29(14)   24.60(10)   33.11(2)      9.83(14)
                       NAYEL                10.13(15)   18.32(15)   11.31(16)     9.76(15)
                       NAYEL                 7.73(16)   24.06(12)   21.07(11)     8.37(16)

Table 13: Full results for Subtask 1.1 (country-level MSA). The numbers in parentheses are the ranks. The table is
sorted on the macro F1 score, the official metric.

                             Team              F1          Acc      Precision     Recall
                       CairoSquad           32.26(1)    51.66(1)    36.03(1)     31.09(1)
                       CairoSquad           31.04(2)    51.02(2)    35.01(2)     30.62(2)
                       CS-UM6P              30.64(3)    49.50(4)    32.91(6)     30.34(3)
                       CS-UM6P              30.14(4)    48.94(5)    33.20(4)     30.21(4)
                       CS-UM6P              29.08(5)    50.30(3)    34.99(3)     29.04(5)
                       IDC team             26.10(6)    42.70(9)    27.04(11)    25.88(6)
                       Phonemer             24.29(7)    44.14(6)    30.24(7)     23.70(7)
                       IDC team             24.00(8)    40.08(14)   25.57(15)    23.29(9)
                       Phonemer             23.56(9)    43.32(8)    28.05(10)    23.34(8)
                       Phonemer             22.72(10)   43.46(7)    28.13(9)     22.55(10)
                       Speech Translation   21.49(11)   40.54(10)   26.75(12)    20.36(12)
                       Arizona              21.37(12)   40.46(12)   26.32(13)    20.78(11)
                       Speech Translation   21.14(13)   40.32(13)   25.43(16)    20.16(14)
                       Speech Translation   21.09(14)   40.50(11)   26.29(14)    20.02(15)
                       Arizona              20.48(15)   40.04(15)   24.09(17)    20.22(13)
                       Arizona              19.85(16)   39.90(16)   22.89(18)    19.66(16)
                       AraDial MJ           18.94(17)   35.94(22)   21.58(22)    18.28(17)
                       NAYEL                18.72(18)   37.16(20)   21.61(21)    18.12(18)
                       AraDial MJ           18.66(19)   35.54(23)   21.45(23)    18.03(19)
                       AraDial MJ           18.09(20)   37.22(19)   21.84(20)    17.55(20)
                       AraDial MJ           18.06(21)   38.48(17)   22.70(19)    17.39(21)
                       IDC team             16.33(22)   29.82(25)   18.04(25)    16.10(22)
                       NAYEL                16.31(23)   38.08(18)   32.94(5)     15.91(23)
                       NAYEL                14.41(24)   32.78(24)   20.16(24)    14.11(24)
                       NAYEL                13.16(25)   36.96(21)   30.00(8)     13.83(25)
                       NAYEL                12.81(26)   26.48(26)   14.32(26)    12.66(26)
                       AraDial MJ            4.34(27)   12.64(27)    4.33(27)     4.70(27)

                          Table 14: Full results for Subtask 1.2 (province-level MSA).
Team          F1         Acc      Precision    Recall
 CairoSquad   6.43(1)     6.66(1)    7.11(1)     6.71(1)
 CairoSquad   5.81(2)     6.24(2)    6.26(2)     6.33(2)
 Phonemer     5.49(3)     6.00(3)    6.17(3)     6.07(3)
 Phonemer     5.43(4)     5.96(4)    6.12(4)     6.02(4)
 CS-UM6P      5.35(5)     5.72(6)    5.71(7)     5.75(6)
 Phonemer     5.30(6)     5.84(5)    5.97(6)     5.90(5)
 CS-UM6P      5.12(7)     5.50(7)    5.24(8)     5.53(7)
 CS-UM6P      4.72(8)     5.00(8)    5.97(5)     5.02(8)
 NAYEL        3.51(9)     3.38(10)   4.09(9)     3.45(10)
 NAYEL        3.47(10)    3.56(9)    3.53(10)    3.60(9)
 NAYEL        3.16(11)    3.28(11)   3.38(12)    3.40(11)
 NAYEL        3.15(12)    3.06(12)   3.43(11)    3.07(12)

Table 15: Full results for Subtask 2.1 (country-level DA).

   Team          F1         Acc      Precision    Recall
 CairoSquad   8.60(1)     9.46(1)    9.07(1)     9.33(1)
 CairoSquad   7.88(2)     8.78(2)    8.27(2)     8.66(2)
 CS-UM6P      7.32(3)     7.92(4)    7.73(4)     7.95(3)
 CS-UM6P      7.29(4)     8.04(3)    8.17(3)     7.90(4)
 CS-UM6P      5.30(5)     6.90(5)    7.00(5)     6.82(5)
 NAYEL        4.55(6)     4.80(10)   4.71(6)     4.55(10)
 NAYEL        4.43(7)     4.88(9)    4.59(8)     4.62(9)
 Phonemer     4.37(8)     5.32(6)    4.49(9)     5.19(6)
 Phonemer     4.33(9)     5.26(7)    4.44(10)    5.14(7)
 Phonemer     4.23(10)    5.20(8)    4.21(11)    5.08(8)
 NAYEL        3.92(11)    4.12(12)   4.05(12)    4.00(12)
 NAYEL        3.02(12)    3.10(13)   3.19(13)    3.19(13)
 CS-UM6P      2.90(13)    4.20(11)   4.68(7)     4.13(11)

Table 16: Full results for Subtask 2.2 (province-level DA).
You can also read